Hallucination is the failure mode of ungrounded generation. The model is built to produce likely text, and likely is not the same as true. When it has no real source to draw on, it fills the gap with plausible invention.
Want to see what a strategy recommendation looks like when you can trace it to its method? See it in a 20-minute demo →
The danger is that the output looks correct. A hallucinated citation reads like a real one. A fabricated figure sits neatly in a sentence. Without grounding and attribution, there is no signal that anything is wrong.
The defence is structural: ground the model in retrievable sources and require every claim to trace back to one.